DeepfakeBench-MM: A Comprehensive Benchmark for Multimodal Deepfake Detection
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: We're looking at the paper: "DeepfakeBench-MM: A Comprehensive Benchmark for Multimodal Deepfake Detection".
Elias: Here's the gist of the paper. DeepfakeBench-MM introduces Mega-MMDF, a large-scale and diverse multimodal deepfake dataset and DeepfakeBench-MM, the first unified benchmark for multimodal deepfake detection The paper presents this infrastructure to address the lack of sufficient and diverse training data and standardized benchmarks hindering deeper exploration Dataset Construction Mega-MMDF is a large-scale multimodal deepfake dataset constructed through an automated deepfake generation framework, featuring three key characteristics: large scale, high forgery diversity, and high quality The dataset contains 0.1 million real samples and 1.1 million forged samples, totaling 1.2 million instances The construction involves generating 21 compositionally diverse forgery pipelines based on combinations of 28 forgery methods across 7 forgery types Quality control is integrated into every pipeline through three assessment modules: audio quality assessment, visual quality assessment, and audio-video synchronization quality assessment Overall dataset quality was validated using objective quantitative metrics like Fréchet Audio Distance (FAD) and Fréchet Video Distance (FVD), alongside human studies involving 60 volunteers Benchmark Development DeepfakeBench-MM is the first unified benchmark for multimodal deepfake detection, supporting 5 multimodal deepfake datasets and 11 detectors within an extensible codebase The benchmark standardizes training/evaluation protocols across detection components and provides a common platform for method analysis Supported datasets include FakeAVCeleb, LAV-DF, AVDeepfake1M, IDForge, and Mega-MMDF Detectors implemented span four categories: Baseline Models, Regular Models (7 detectors), Ensemble Models (one model), and Pretrained Multimodal Large Language Models (two models) The evaluation protocol focuses on intra-dataset performance, cross-dataset generalization, and cross-pipeline generalization using the AUC metric Analysis of Findings Analysis 1 observed that although detectors generally achieve high AUCs within Mega-MMDF (up to 0.994), their performance drops markedly when applied to external datasets Analysis 2 revealed that in stacked forgeries, artifacts contribute unequally, following the hierarchy EFS > FS > FR Analysis 3 showed that finetuning the second-stage backbones of detectors like AVTS and AVFF yields consistent improvements across datasets Analysis 4 indicated a consistent bias toward the visual modality when detectors operate in a single-modality scenario, suggesting that late-fusion architectures often lead to modality competition rather than complementarity Future Directions The key takeaway messages suggest that current multimodal detectors rely more on visual artifacts and need new architectures and fusion strategies to mitigate modality bias Future work should explore finetuning the second-stage backbones or randomly masking the dominant modality during training to achieve more balanced results Mitigating artifact bias in visual modalities is key to learning richer and more diverse forgery cues The paper concludes that Mega-MMDF and DeepfakeBench-MM provide a solid foundation for future research in multimodal deepfake detection In this work, the authors first introduce Mega-MMDF, a million-scale, diverse, and high-quality dataset designed to advance research in multimodal deepfake detection Building upon it, they develop DeepfakeBench-MM, the first unified benchmark that standardizes preprocessing and evaluation protocols for multimodal deepfake detection Leveraging these infrastructures allows for fair comparisons and facilitates the development of new methods The release of Mega-MMDF and DeepfakeBench-MM is believed to provide a solid foundation for future research in this area References Enes Altuncu, Virginia NL Franqueira, and Shujun Li.
Nadia: Deepfake: Definitions, performance metrics and standards, datasets, and a meta-review.
Elias: Frontiers in Big Data, 2024 Junyi Ao et al.
Nadia: Speecht5: Unified-modal encoder-decoder pre-training for spoken language processing.
Paper summary: Nadia: It’s about providing the infrastructure needed so we can actually compare different detection methods fairly against one another, which is a big step forward for how we approach defense against deepfakes in practice.
Elias: The authors are addressing the lack of sufficient and diverse training data, and they've built a comprehensive benchmark that tackles both data scarcity and evaluation standardization at once.
Priya: From my side, I think the real value here is moving away from detectors that only work well on one specific type of forgery artifact toward systems that can handle the full spectrum of multimodal fakes we see in the wild.
Nadia: Yeah, so we have this massive dataset and a standardized testbed to push the limits on what multimodal deepfake detection can actually achieve when it has to deal with real-world complexity.
Conclusion: Nadia: So we’ve seen how they built this huge set of real and fake data, and now they have a way to test all these different detectors against it all at once.
Elias: They call it DeepfakeBench-MM, and honestly, the authors are focused on making sure everyone is using the same rules for training and testing these detection models.
Priya: What I see here is that by standardizing the setup, they’re trying to get past just seeing if one detector works well on one dataset.
Nadia: Right, so instead of just saying a model got a high score on Mega-MMDF and then hoping it works elsewhere, they are testing how well it generalizes across different kinds of fakes and even different types of data.
Elias: They’ve included five different deepfake datasets and eleven detectors to see if those models can actually handle the variety.
Priya: That standardization is important because it forces us to look at where those models fail, not just where they succeed on a neat little test set.
Nadia: It really comes down to whether these tools are robust enough for real-world use against the messy stuff people are actually making now.
Elias: The authors have built a solid foundation here, and it gives us a common place to start looking at how well our detection strategies hold up under pressure.
Priya: And looking at the findings, they point out that there’s still this bias towards visual details in some single-modality setups, which tells us we need better ways to balance what the system is actually paying attention to.
Nadia: Yeah, so it suggests that future work needs to focus on making those models less reliant on just the visual stuff when they are dealing with audio too.
Elias: Exactly, and that’s where we go next because understanding this limitation tells us exactly where we need to push our research in the next few years.
The Chinese University of Hong Kong, Shenzhen · University at Buffalo, State University of New York · Nanjing University of Posts and Telecommunications
cs.CR, cs.CV, cs.MM
Submitted: 2025-10-26
Updated: 2026-11-14
Comments: Preprint
Code: https://github.com/haofanwang/inswapper
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 98/100
The gist: DeepfakeBench-MM introduces Mega-MMDF, a large-scale and diverse multimodal deepfake dataset and DeepfakeBench-MM…
Key concepts
- Mega-MMDF
- A massive, high-quality dataset containing 1.2 million real and forged samples. It was created using an automated framework that generated 21 compositionally diverse forgery pipelines based on various methods and types, ensuring large scale and high diversity.
- DeepfakeBench-MM
- The first unified benchmark for multimodal deepfake detection. It supports five different multimodal datasets and eleven distinct detectors within an extensible codebase. It standardizes protocols for training and evaluation, allowing for fair comparisons across different methods.
- Multimodal Deepfake Detection
- The field of detecting manipulated media that combines visual (video) and auditory (audio) elements. This paper focuses on creating tools and benchmarks to accurately identify deepfakes where both the video content and the accompanying sound are intentionally forged or manipulated.
- Modality Competition
- A phenomenon observed in single-modality detectors where they show a consistent bias toward the visual modality. This suggests that late-fusion architectures might lead to competition between modalities rather than complementary information when detecting deepfakes.
Terminology
Summary
DeepfakeBench-MM introduces Mega-MMDF, a large-scale and diverse multimodal deepfake dataset and DeepfakeBench-MM, the first unified benchmark for multimodal deepfake detection The paper presents this infrastructure to address the lack of sufficient and diverse training data and standardized benchmarks hindering deeper exploration
Dataset Construction
Mega-MMDF is a large-scale multimodal deepfake dataset constructed through an automated deepfake generation framework, featuring three key characteristics: large scale, high forgery diversity, and high quality The dataset contains 0.1 million real samples and 1. p. 0 Mega-MMDF currently
contains 0. 1 million real samples and 1. 1 million real samples and 1. 1 million forged samples, making it one of
the largest and most diverse multimodal deepfake datasets, with plans for continuous expansion.Where the paper says it
Where the paper says it
Benchmark Development
DeepfakeBench-MM is the first unified benchmark for multimodal deepfake detection, supporting 5 multimodal deepfake datasets and 11 detectors within an extensible codebase The benchmark standardizes training/evaluation protocols across detection components and provides a common platform for method analysis Supported datasets include FakeAVCeleb, LAV-DF, AVDeepfake1M, IDForge, and Mega-MMDF Detectors implemented span four categories: Baseline Models, Regular Models (7 detectors), Ensemble Models (one model), and Pretrained Multimodal Large Language Models (two models) The evaluation protocol focuses on intra-dataset performance, cross-dataset generalization, and cross-pipeline generalization using the AUC metric
Analysis of Findings
Analysis 1 observed that although detectors generally achieve high AUCs within Mega-MMDF (up to 0.994), their performance drops markedly when applied to external datasets Analysis 2 revealed that in stacked forgeries, artifacts contribute unequally, following the hierarchy EFS > FS > FR Analysis 3 showed that finetuning the second-stage backbones of detectors like AVTS and AVFF yields consistent improvements across datasets Analysis 4 indicated a consistent bias toward the visual modality when detectors operate in a single-modality scenario, suggesting that late-fusion architectures often lead to modality competition rather than complementarity
Future Directions
The key takeaway messages suggest that current multimodal detectors rely more on visual artifacts and need new architectures and fusion strategies to mitigate modality bias Future work should explore finetuning the second-stage backbones or randomly masking the dominant modality during training to achieve more balanced results Mitigating artifact bias in visual modalities is key to learning richer and more diverse forgery cues The paper concludes that Mega-MMDF and DeepfakeBench-MM provide a solid foundation for future research in multimodal deepfake detection
Conclusion
In this work, the authors first introduce Mega-MMDF, a million-scale, diverse, and high-quality dataset designed to advance research in multimodal deepfake detection Building upon it, they develop DeepfakeBench-MM, the first unified benchmark that standardizes preprocessing and evaluation protocols for multimodal deepfake detection Leveraging these infrastructures allows for fair comparisons and facilitates the development of new methods The release of Mega-MMDF and DeepfakeBench-MM is believed to provide a solid foundation for future research in this area
References
Enes Altuncu, Virginia NL Franqueira, and Shujun Li. Deepfake: Definitions, performance metrics and standards, datasets, and a meta-review. p. 9 REFERENCES
Enes Altuncu, Virginia NL Franqueira, and Shujun Li. Deepfake: Definitions, performance metrics
and standards, datasets, and a meta-review. Frontiers in Big Data, 2024. Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing
Li, Yu Zhang, et al. Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing
Li, Yu Zhang, et al. Speecht5: Unified-modal encoder-decoder pre-training for spoken language
processing. In Association for Computational Linguistics, 2022. Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Whisperx: Time-accurate speech
transcription of long-form audio. Interspeech, 2023. Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, and Munawar Hayat. Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, and Munawar Hayat. Glitch
in the matrix: A large scale benchmark for content driven audio–visual forgery detection and
localization. Computer Vision and Image Understanding, 2023. Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, and
Kalin Stefanov. Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, and
Kalin Stefanov. Av-deepfake1m: A large-scale llm-driven audio-visual deepfake dataset. In ACM
International Conference on Multimedia, 2024. Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. Edresson Casanova, Julian Weber, Christopher D Shulby, Arnaldo Candido Junior, Eren Gölge, and
Moacir A Ponti. Edresson Casanova, Julian Weber, Christopher D Shulby, Arnaldo Candido Junior, Eren Gölge, and
Moacir A Ponti. Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for
everyone. Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for
everyone. In International Conference on Machine Learning, 2022. Abhra Chaudhuri, Anjan Dutta, Tu Bui, and Serban Georgescu. A closer look at multimodal
representation collapse. In International Conference on Machine Learning, 2025. Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi
Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi
Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal
modeling and audio understanding in video-llms. arXiv e-prints, 2024. Ha-Yeong Choi, Sang-Hoon Lee, and Seong-Whan Lee. Ha-Yeong Choi, Sang-Hoon Lee, and Seong-Whan Lee. Diff-hiervc: Diffusion-based hierarchical
voice conversion with robust pitch generation and masked prior for zero-shot speaker adaptation. Whisperx: Time-accurate speech
transcription of long-form audio. Interspeech, 2023. Ha-Yeong Choi, Sang-Hoon Lee, and Seong-Whan Lee. Diff-hiervc: Diffusion-based hierarchical
voice conversion with robust pitch generation and masked prior for zero-shot speaker adaptation. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. Building on it, we present DeepfakeBench-MM, the first unified
benchmark for multimodal deepfake detection. It establishes standardized protocols
across the entire detection pipeline and serves as a versatile platform for evaluating
existing methods as well as exploring novel approaches. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Whisperx: Time-accurate speech
transcription of long-form audio. Av-deepfake1m: A large-scale llm-driven audio-visual deepfake dataset. In ACM
International Conference on Multimedia, 2024. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. , 2025), in which
multimodal models fail to leverage both modalities in a balanced manner. Since most detectors adopt a
late-fusion architecture, the two modalities often act in competition rather than complementarity. Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for
everyone. In International Conference on Machine Learning, 2022. , 2024), we consider
three common types of visual-only forgeries: Face Swapping (FS), Entire Face Synthesis (EFS),
and Face Editing (FE). FS swaps the facial ID of the target video with the ID of the source video,
while preserving all remaining contents (e. Av-deepfake1m: A large-scale llm-driven audio-visual deepfake dataset. In ACM
International Conference on Multimedia, 2024. For visual forgery methods that output static images, we employ audio-driven face reenactment (FR)
to generate temporally consistent videos that align with the audio. Similarly, for video-based visual
forgeries where the manipulated video does not naturally synchronize with the forged audio, we
apply face reenactment to enforce cross-modal consistency, mimicking real-world deepfake scenarios. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. 1 REAL DATA COLLECTION
Our real-world audiovisual samples are drawn from three public datasets: Voxceleb2 (Chung et al. ,
2018), MEAD (Wang et al. To address
this challenge, we first build Mega-MMDF, a large-scale, diverse, and high-quality
dataset for multimodal deepfake detection. Specifically, we employ 21 forgery
pipelines through the combination of 10 audio forgery methods, 12 visual forgery
methods, and 6 audio-driven face reenactment methods. ,
2018), MEAD (Wang et al. , 2020a), and CelebV-HQ (Zhu et al. 4 OVERALL QUALITY
After the dataset construction process, we validate overall dataset quality through (1) Objective
Quantitative Metrics using Fréchet Audio Distance (FAD) (Kilgour et al. , 2019) and Fréchet
Video Distance (FVD) (Unterthiner et al. ,
2020b; Wu et al. , 2015), we consider
three common types of audio forgeries: Text-to-Speech (TTS), Voice
Conversion (VC), and Partial Fake
(PF). This
architecture can lead to predictive cues from one modality to become entangled with noise features
from the other, suppressing informative signals (Chaudhuri et al. While such modality bias
has been extensively discussed in multimodal representation learning, it remains largely unaddressed
in current multimodal deepfake detection frameworks, highlighting a critical design limitation that
suggests future detectors should incorporate strategies, such as bias-aware training or adaptive
fusion, to mitigate modality dominance and fully exploit the strengths of each modality. Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for
everyone. In International Conference on Machine Learning, 2022. , 2015), we consider
three common types of audio forgeries: Text-to-Speech (TTS), Voice
Conversion (VC), and Partial Fake
(PF). Given a text input and reference audio, TTS generates fake audio adhering to the speaker’s tone of
the reference audio. ,
2020b; Wu et al. , 2015), we consider
three common types of audio forgeries: Text-to-Speech (TTS), Voice
Conversion (VC), and Partial Fake
(PF). To address
this challenge, we first build Mega-MMDF, a large-scale, diverse, and high-quality
dataset for multimodal deepfake detection. Specifically, we employ 21 forgery
pipelines through the combination of 10 audio forgery methods, 12 visual forgery
methods, and 6 audio-driven face reenactment methods. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. For consistency with prior work (Yan et al. ,
2023; 2024), we report results using the widely adopted AUC metric. Videollama 2: Advancing spatial-temporal
modeling and audio understanding in video-llms. arXiv e-prints, 2024. Av-deepfake1m: A large-scale llm-driven audio-visual deepfake dataset. In ACM
International Conference on Multimedia, 2024. Av-deepfake1m: A large-scale llm-driven audio-visual deepfake dataset. In ACM
International Conference on Multimedia, 2024. However, the lack of sufficient and diverse training data, along with
the absence of a standardized benchmark, hinder deeper exploration. To address
this challenge, we first build Mega-MMDF, a large-scale, diverse, and high-quality
dataset for multimodal deepfake detection. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. , 2018) to measure audio/video fidelity by quantifying
real-forgery sample distribution gaps, and to evaluate synchronization using Sync-C and Sync-D
calculated via a pretrained SyncNet (Chung & Zisserman, 2017) model. (2) Human Studies involving
60 volunteers from diverse backgrounds (e. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. , 2024; Altuncu et al. , 2024) have provided comprehensive overviews of single-modality
datasets, spanning image-, audio-, and video-based forgeries. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. Edresson Casanova, Julian Weber, Christopher D Shulby, Arnaldo Candido Junior, Eren Gölge, and
Moacir A Ponti. Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for
everyone. In International Conference on Machine Learning, 2025. Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi
Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Several multimodal deepfake detectors, such as AVFF and AVTS, adopt a two-phase paradigm. Backbones are first pretrained on large-scale real audio-video pairs to learn cross-modal synchronization,
and then trained on deepfake datasets for classification. Several multimodal deepfake detectors, such as AVFF and AVTS, adopt a two-phase paradigm. Backbones are first pretrained on large-scale real audio-video pairs to learn cross-modal synchronization,
and then trained on deepfake datasets for classification. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. To address
this challenge, we first build Mega-MMDF, a large-scale, diverse, and high-quality
dataset for multimodal deepfake detection. Specifically, we employ 21 forgery
pipelines through the combination of 10 audio forgery methods, 12 visual forgery
methods, and 6 audio-driven face reenactment methods. , 2023), a speech-to-text (STT) model. (b)
SNR quantifies the ratio of meaningful signal to background noise. , 2024; 2023; Cai et al. , 2024), we consider
three common types of visual-only forgeries: Face Swapping (FS), Entire Face Synthesis (EFS),
and Face Editing (FE). Glitch
in the matrix: A large scale benchmark for content driven audio–visual forgery detection and
localization. Computer Vision and Image Understanding, 2023. The dominance of EFS in certain combinations reflects
an artifact bias phenomenon, where specific artifacts disproportionately influence the learned
representation. This bias can skew cross-pipeline generalization and highlights the need for future
work on representation balancing to ensure fairer exploitation of multi-artifact information. A closer look at multimodal
representation collapse. In International Conference on Machine Learning, 2025. This
architecture can lead to predictive cues from one modality to become entangled with noise features
from the other, suppressing informative signals (Chaudhuri et al. While such modality bias
has been extensively discussed in multimodal representation learning, it remains largely unaddressed
in current multimodal deepfake detection frameworks, highlighting a critical design limitation that
suggests future detectors should incorporate strategies, such as bias-aware training or adaptive
fusion, to mitigate modality dominance and fully exploit the strengths of each modality. Glitch
in the matrix: A large scale benchmark for content driven audio–visual forgery detection and
localization. Computer Vision and Image Understanding, 2023. This
architecture can lead to predictive cues from one modality to become entangled with noise features
from the other, suppressing informative signals (Chaudhuri et al. While such modality bias
has been extensively discussed in multimodal representation learning, it remains largely unaddressed
in current multimodal deepfake detection frameworks, highlighting a critical design limitation that
suggests future detectors should incorporate strategies, such as bias-aware training or adaptive
fusion, to mitigate modality dominance and fully exploit the strengths of each modality. Videollama 2: Advancing spatial-temporal
modeling and audio understanding in video-llms. arXiv e-prints, 2024. Preprint
DEEPFAKEBENCH-MM: A COMPREHENSIVE BENCHMARK FOR MULTIMODAL DEEPFAKE DETECTION
Kangran Zhao1∗
, Yupeng Chen1∗
, Xiaoyu Zhang1∗
, Yize Chen1
, Weinan Guan1
Baicheng Chen1
, Chengzhe Sun2
, Soumyya Kanti Datta2
Qingshan Liu3
, Siwei Lyu2
, Baoyuan Wu1†
1The Chinese University of Hong Kong, Shenzhen
2University at Buffalo, State University of New York
3Nanjing University of Posts and Telecommunications
ABSTRACT
The misuse of advanced generative AI models has resulted in the widespread proliferation of falsified data, particularly forged human-centric audiovisual content,
which poses substantial societal risks (e. , financial fraud and social instability). Videollama 2: Advancing spatial-temporal
modeling and audio understanding in video-llms. arXiv e-prints, 2024. , 2024),
which contains 860,039 fake samples, the scale of fake data in other datasets remains relatively small. (2) Low forgery diversity: Most datasets are generated using only a few forgery methods (e. Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for
everyone. In International Conference on Machine Learning, 2022. , 2024; 2022; Li et al. , 2023), facial images/videos (Shiohara et al. , 2024; 2023; Cai et al. , 2024), we consider
three common types of visual-only forgeries: Face Swapping (FS), Entire Face Synthesis (EFS),
and Face Editing (FE). In International Conference on Machine Learning, 2025. Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi
Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. FRADE (Nie et al. , 2024) introduces an AFI adapter to
improve attention to high-frequency details and an ACI module to fuse audio and visual features. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. Details of the 28 forgery methods and the 21
forgery pipelines are provided in Appendix A. 3 QUALITY CONTROL MECHANISM DURING DEEPFAKE DATA CONSTRUCTION
To ensure the quality of the constructed forgery data, we design a quality control mechanism
comprising three assessment modules: audio quality assessment, visual quality assessment, and
audio-video synchronization quality assessment. , 2024) to assess the lip synchronization
quality of manipulated audiovisual data. 4 OVERALL QUALITY
After the dataset construction process, we validate overall dataset quality through (1) Objective
Quantitative Metrics using Fréchet Audio Distance (FAD) (Kilgour et al. Glitch
in the matrix: A large scale benchmark for content driven audio–visual forgery detection and
localization. Computer Vision and Image Understanding, 2023. To address
this challenge, we first build Mega-MMDF, a large-scale, diverse, and high-quality
dataset for multimodal deepfake detection. Specifically, we employ 21 forgery
pipelines through the combination of 10 audio forgery methods, 12 visual forgery
methods, and 6 audio-driven face reenactment methods. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. This
architecture can lead to predictive cues from one modality to become entangled with noise features
from the other, suppressing informative signals (Chaudhuri et al. While such modality bias
has been extensively discussed in multimodal representation learning, it remains largely unaddressed
in current multimodal deepfake detection frameworks, highlighting a critical design limitation that
suggests future detectors should incorporate strategies, such as bias-aware training or adaptive
fusion, to mitigate modality dominance and fully exploit the strengths of each modality. Av-deepfake1m: A large-scale llm-driven audio-visual deepfake dataset. In ACM
International Conference on Multimedia, 2024. However, the lack of sufficient and diverse training data, along with
the absence of a standardized benchmark, hinder deeper exploration. To address
this challenge, we first build Mega-MMDF, a large-scale, diverse, and high-quality
dataset for multimodal deepfake detection. Glitch
in the matrix: A large scale benchmark for content driven audio–visual forgery detection and
localization. Computer Vision and Image Understanding, 2023. (2) Low forgery diversity: Most datasets are generated using only a few forgery methods (e. ,
AVDeepfake1M uses three while FakeAVCeleb uses four), making detectors prone to overfitting to
specific forgery artifacts and failing to generalize to the wide variety of forgeries encountered in the
wild. Glitch
in the matrix: A large scale benchmark for content driven audio–visual forgery detection and
localization. Computer Vision and Image Understanding, 2023. By
contrast, AVH benefits from pretrained lip-reading checkpoints, enabling it to exploit cross-modal
consistency and achieve superior performance in detecting partial manipulations. Moreover, we note
that pretrained large models, when evaluated in a zero-shot setting, obtain the lowest AUC scores on
Mega-MMDF, highlighting both the difficulty and the novelty of our dataset. Videollama 2: Advancing spatial-temporal
modeling and audio understanding in video-llms. arXiv e-prints, 2024. Edresson Casanova, Julian Weber, Christopher D Shulby, Arnaldo Candido Junior, Eren Gölge, and
Moacir A Ponti. Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for
everyone. In International Conference on Machine Learning, 2025. Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi
Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. , 2023; Lee et al. , 2024) to assess the lip synchronization
quality of manipulated audiovisual data. Videollama 2: Advancing spatial-temporal
modeling and audio understanding in video-llms. arXiv e-prints, 2024. End-to-end
reconstruction-classification learning for face forgery detection. In Computer Vision and Pattern
Recognition, 2022. This approach can be regarded as a bias-mitigation form of data augmentation, which is rarely
explored in multimodal forgery detection. Given that models trained on Mega-MMDF generally
exhibit a strong preference for the visual modality, we randomly mask the video stream during
training and supervise using audio-modality labels.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Augment multimodal deepfake detectors with finetuning of second-stage backbones to adapt to forgery-specific artifacts, as shown by
finetuning yields consistent improvements across datasets
in Table 3. This allows models like AVTS and AVFF to overcome the limitations of frozen backbones by adapting backbone representations to artifact-specific cues. -
Implement a training strategy involving random modality masking during training to encourage better exploitation of weaker modalities, as suggested by
randomly masking the video stream during training and supervise using audio-modality labels
for AVTS and MRDF. This aims to mitigate modality dominance in models that show aconsistent bias toward the visual modality.
-
Design fusion strategies that explicitly address modality competition rather than simple averaging, aiming to mitigate the
modalities often act in competition rather than complementarity
issue, as highlighted by the analysis of late-fusion architectures. This could involve developing adaptive fusion methods to ensure complementary use of all modalities. -
Develop a hierarchical or artifact-aware representation learning approach that accounts for the hierarchy of forgery artifacts, such as EFS > FS > FR, to improve cross-pipeline generalization across diverse manipulations. This addresses the finding that
EFS+FR generalizes much better to other EFS-related pipelines than FS-only pipelines.
-
Implement a standardized preprocessing pipeline within the benchmark codebase that incorporates hybrid face detection strategies, specifically discarding frames with no detected face or multiple faces, to ensure clear speaker-face correspondence. This standardizes data preparation across heterogeneous datasets like Mega-MMDF and others, preventing issues like
background region cropped sometimes.
-
Utilize pretrained multimodal large language models like Qwen2.5-Omni or VideoLLaMA2 for zero-shot detection by leveraging their
robust cross-modal representation and generation capabilities.
This allows for immediate evaluation of zero-shot performance on unseen deepfake content without requiring extensive fine-tuning on specific forgery types. -
Create an ensemble model that explicitly combines the results from visual and audio detectors using modality-specific backbones, ensuring that each detector focuses on its modality's strengths before combining predictions with an
averaging strategy.
This serves as a competitive baseline by exploitingcomplementary information across modalities.
Abstract
Audio-visual deepfakes can spread misinformation and enable impersonation, undermining trust in digital media. However, existing datasets provide limited coverage of forgery operations and their combinations, constraining detector development and systematic evaluation. We introduce Mega-MMDF, containing 100,000 real and 1.2 million forged clips across 22 pipelines spanning modality-specific manipulations, stacked forgeries, and joint audio-video generation. To assess its training value and enable fair comparisons, we develop DeepfakeBench-MM, which standardizes training and evaluation across 8 datasets and 10 detectors. Training-based detectors achieve their highest mean AUC when trained on Mega-MMDF, with gains of up to 0.0784 over the best alternative training dataset for each detector. Experiments on data scale and pipeline coverage further support the value of large, diverse training data. Cross-pipeline evaluations reveal bidirectional and asymmetric generalization. Training on composite forgeries does not necessarily enable detection of their constituent operations in isolation. Our analysis of altered inputs shows that modality masking can increase false alarms and missed detections, and temporal misalignment can raise mean predicted forgery probabilities. Together, these resources provide a foundation for understanding and improving multimodal deepfake detection, with publicly available code and dataset access subject to application and usage terms.
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data